Visual Target Detection And Tracking

Researches on computer vision have been booming since 1980's, when Maar proposed the framework of the visual system. A lot of new fields have emerged, such as Image Segmentation, Visual Motion Analysis, 3D reconstruction, Modeling and Recognitions, stereovision and etc. This thesis focuses on the study of visual target detection.Visual target detection and tracking contains two aspects. (1) Object's detection, feature analysis, state and behavior understanding in image sequences; (2) Object's localization and feature extraction in a single image.Upon the human object detected in video, face region will be localized through face detection. Then, all the facial features are extracted by feature pre-position and adjusting. The features' comparing gets human body's classification and identification.Tracing in the image sequences includes temporal difference analysis, object localization, feature and state analysis to understand the behaviors of objects. Besides, the next state of objects can be predicted.The main process of object detection in a single image is localizing the target in the image. Given some image templates, the features of objects in them are extracted, such as textures and profiles. When new image comes, the detection window is defined, and features in the widow are extracted to compare with those templates.This paper introduces technology development of visual object detection and tracking, including face detection, facial feature detection, video object tracking. We also propose a real-time moving target tracking and identification algorithm. Moving object's information is obtained by temporal difference analysis. Through segmentation and counter extraction, we get standardization vector. Object's shape information has been classified by PCA. The next state forecast through the statistics of the state transferring. Face feature detection is present in this paper. The searching scope is narrowed by using face diction. The scope is further strengthening restricted by features' fixed ratio information. By transferring, scaling, rotating the initial localization of the feature points, we make feature point positioning approximate expectations. At last step, we adjust location details through the gradient vector.